Xiaoliu BOT
中文EN

X Platform July 11 AI Brief | Apple Sues OpenAI for Theft, GPT-5.6 Sol Raises Safety Concerns, AI Video Workflows Emerge

Apple Formally Sues OpenAI for Systematic Commercial Theft

Apple has filed a lawsuit against OpenAI in California federal court, accusing it of systematically stealing Apple’s trade secrets for AI hardware development. The defendants include OpenAI, hardware head Tang Tan (former Apple iPhone/Apple Watch product design VP with 24 years at Apple), former Apple senior systems electrical engineer Chang Liu, and io Products co-founded by Jony Ive. Chang Liu left Apple in January 2026 to join OpenAI but secretly kept an Apple-issued computer, using a vulnerability from OpenAI’s office to re-login to Apple’s private servers and download over 1,000 pages of confidential engineering documents within weeks. Apple’s complaint states that OpenAI has hired over 400 former Apple engineers and executives, and its hardware interview process requires candidates to prepare a “technical deep dive” PPT and ask questions using Apple’s confidential project codenames. OpenAI is also accused of contacting Apple’s exclusive industrial design partners to demand the execution of Apple’s proprietary metal finishing processes. Apple believes OpenAI’s hardware business “fundamentally fails due to its illegal reliance on stolen trade secrets.”

Sources:

GPT-5.6 Sol Deep Review: Great Daily Assistant, Fable Still the Ultimate Solution

The Every team published a comprehensive review of GPT-5.6 Sol, concluding that Sol is an excellent daily collaboration model, but Claude Fable remains stronger for the most complex tasks. Sol’s strengths include speed, stability, instruction following, and proactive context seeking (it can read 46 CSVs, identify missing information, and return with suggestions). Its execution and persistence significantly outperform GPT-5.5. However, in Every’s “senior engineer” benchmark, Fable scored 90/100 while Sol scored only 56/100—Sol’s weakness lies in not knowing when to stop, adding approximately 12,900 lines of code in a rewrite task, where each change is reasonable individually but overall overly complex. Dan Shipper likened Sol to a Porsche and Fable to a warp drive. Many engineers’ practical approach is to use Fable as the orchestrator to plan and Sol for execution. The Codex team’s AMA also provided Effort selection strategies: daily use Sol Medium, lightweight Terra, ambiguous bugs High, and only production incidents Ultra.

Sources:

GPT-5.6 Sol Accidentally Deletes User Files, Sparking Widespread Concern

Matt Shumer, former CEO of HyperWrite and well-known AI investor, posted that GPT-5.6-Sol accidentally deleted almost all files on his Mac. Details: while testing the Ultra mode invited by OpenAI, a sub-agent responsible for code review executed a cleanup command, misinterpreting $HOME as the entire user directory and directly running rm -rf /Users/mattsdevbox. Ironically, the culprit was the review sub-agent meant to catch errors. Multiple bloggers shared this incident and compared it to Claude Fable, noting that Fable is more trustworthy in sandbox isolation and security. This event also sparked discussions about the safety of running coding agents locally.

Sources:

ChatGPT Work / Codex Major Updates and Product Restructuring

OpenAI launched ChatGPT Work agent mode, unifying the desktop client by merging Chat/Work/Codex modes into a single app. Work can collect information across applications to deliver finished documents, spreadsheets, or web apps, while Codex focuses on code repository operations. The OpenAI developer relations team announced Build Week (July 14) and will host a livestream on July 13. OpenAI product lead Tibo stated that all user usage quotas have been reset and will be reset again, acknowledging that the release’s default settings easily lead users to unconsciously consume high-compute modes, the desktop restructuring was too aggressive, and some multi-agent workflows have degraded. Anthropic also made corresponding moves: Claude Code desktop added a built-in browser, the desktop app changed from three tabs to Home/Code tabs, and Cowork is expanding to web and mobile with cross-device task synchronization.

Sources:

Grok 4.5 Free to All, Speed Widely Praised

Grok officially announced that Grok 4.5 is now available to the free tier via Grok Build. Multiple bloggers gave positive reviews: some consider it among the top three models globally, matching Opus 4.8 on most tasks; others praised its output efficiency, calling it “a SpaceX rocket-like experience.” xAI’s low-cost, high-efficiency approach is rapidly gaining user approval, especially in speed and cost-effectiveness.

Sources:

LangChain Releases OpenWiki 0.1.0, Fable’s Video Editing Capabilities Spark Discussion

LangChain launched OpenWiki 0.1.0, adding proactive memory for agents: it automatically fetches data from six sources—Gmail, Notion, Git repositories, X/Twitter, Hacker News, and web search—organizes them into a local Markdown wiki for agents to call anytime, and auto-refreshes on a set schedule without manual maintenance. A video editing demo by Japanese blogger paji_a using Fable garnered significant attention, with many users sharing it and calling the results “scary”—it can automatically add subtitles, adjust colors and fonts, and is considered a game-changer for video post-production. Anthropic further extended Fable’s trial access.

Sources:

Practical AI Video and Image Creation Workflows Continue to Emerge

Multiple creators shared high-quality AI video/image generation methods. Derek_wall90176 released a complete workflow using “12-grid storyboard + Seedance” to generate Shaolin Nanquan martial arts moves, solving the pain point of AI-generated martial arts lacking force feedback and logical sequences. Gui Zang used GPT-5.6 Sol to create a “Life Achievement Generator” web tool in half a day. Xingzhe AI (joshesye) coded a short drama tool called “Dream” during a meeting using Codex. Xiaoyu Chengzi used GPT Image 2.0 + Seedance 2.0 mini to create a One Piece character three-view outfit video template. Some bloggers also noted that Chinese state media have started using GPT to generate images.

Sources:

Statistics: Scan timeline entries=360 Bloggers hit=31 Total tweets hit=167 Weighted tweet score=127.4 Original tweets=75 RT tweets=41 Fetch attempts=2 Boundary coverage status=tail_confidently_crossed_target_boundary